deep_lincs

围绕LINCS L1000表达数据的Keras深度学习包装。

请在此处查看文档 这里。

安装

$ pip install deep-lincs

入门

1.3百万个L1000配置文件的数据在GEO上可用。脚本load_files.sh获取所有可用的Level 3数据和元数据。最大的文件相当大（约50Gb），请耐心等待。

$ git clone https://github.com/manzt/deep_lincs.git && cd deep_lincs
$ source load_files.sh # download raw data from GEO
$ cd notebooks 
$ jupyter lab # get started in a notebook 

L1000数据集

Dataset类通过多种方法构建，用于加载、子集、过滤和组合表达和元数据。

from deep_lincs import Dataset

# Select samples 
cell_ids = ["VCAP", "MCF7", "PC3"]
pert_types = ["trt_cp", "ctl_vehicle", "ctl_untrt"]

# Loading a Dataset
dataset = Dataset.from_yaml("settings.yaml", cell_id=cell_ids, pert_type=pert_types)

# Normalizing the expression data
dataset.normalize_by_gene("standard_scale")

# Chainable methods
subset = dataset.sample_rows(5000).filter_rows(pert_id=["ctl_vehicle", "ctl_untrt"])

模型

模型通过Dataset类与接口简单地进行不同架构的训练和评估。

单个分类器

from deep_lincs.models import SingleClassifier

model = SingleClassifier(dataset, target="cell_id")
model.prepare_tf_datasets(batch_size=64)
model.compile_model([128, 128, 64, 32], dropout_rate=0.1)
model.fit(epochs=10)

model.evaluate() # Evaluates on isntance test Dataset
model.evaluate(subset) # Evalutates model on user-defined Dataset

多个分类器

from deep_lincs.models import MutliClassifier

targets = ["cell_id", "pert_type"]
model = MutliClassifier(dataset, target=targets)
model.prepare_tf_datasets(batch_size=64)
model.compile_model(hidden_layers=[128, 128, 64, 32])
model.fit(epochs=10)

model.evaluate() # Evaluates on isntance test Dataset
model.evaluate(subset) # Evalutates model on user-defined Dataset

自编码器

from deep_lincs.models import AutoEncoder

model = AutoEncoder(dataset)
model.prepare_tf_datasets(batch_size=64)
model.compile_model(hidden_layers=[128, 32, 128], l1_reg=0.01)
model.fit(epochs=10)

model.encoder.predict() # Gives encodings for instance test Dataset
model.encoder.predict(subset) # Gives encodings for user-defined Dataset